Cannot use MCP Optimizer with Podman Desktop on Windows

[danbarr]

Podman Desktop on macOS works well with the MCP Optimizer. But on Windows, the optimizer cannot find the API server during its port scan on host.docker.internal. That host exists in Podman Desktop, but on Windows, it resolves to the WSL Linux host that the podman engine runs under, not the Windows host which is where the thv API server is listening.

This is different than the behavior with Docker Desktop in Windows, which DOES return the Windows host for host.docker.internal.

There's a longstanding issue in the podman repo about this: containers/podman#14933

• OS: Windows 11
• Container engine: Podman Desktop v1.22.1, Podman engine v5.6.2
• ToolHive version: v0.11.1
• MCP Optimizer version: 0.0.4

[aponcedeleonch]

This is essentially a networking issue of Podman + Windows as noted by Dan. The broken part is Optimizer trying to reach ToolHive server on the host machine. However, the problem is applicable for any Podman container in Windows trying to reach a server running in the host. Some troubleshooting was done to check if there was a way to get around it.

Test setup

1. Create a simple http server in Powershell. This will act as the ToolHive server

@"
`$listener = New-Object System.Net.HttpListener
`$listener.Prefixes.Add('http://localhost:8000/')
`$listener.Start()
Write-Host 'Server running on http://localhost:8000/ - Press Ctrl+C to stop'
while (`$listener.IsListening) {
    `$context = `$listener.GetContext()
    `$response = `$context.Response
    `$content = [System.Text.Encoding]::UTF8.GetBytes('Hello from host!')
    `$response.ContentLength64 = `$content.Length
    `$response.OutputStream.Write(`$content, 0, `$content.Length)
    `$response.Close()
}
"@ | Out-File -FilePath server.ps1

# Run it
powershell -File server.ps1

2. Spin up test containers. The test containers act as MCP Optimizer and some other MCP server

# Container 1: Running nginx on port 8081
podman run -d --name container1 -p 8081:80 nginx:alpine

# Container 2: Running nginx on port 8082
podman run -d --name container2 -p 8082:80 nginx:alpine

Tests performed

1. Trying to reach host from container using $(hostname).local

# Go into container 1
podman exec -it container1 sh

# Try to reach server on host <--- Failed
curl http://$(hostname).local:8000

# Try to reach server on host <--- Failed
curl http://host.containers.internal:8000

# Try to reach server on host <--- Failed
curl http://localhost:8000

# Try to reach the other container <--- Worked
curl http://host.containers.internal:8082

2. Trying to add --add-host=host.containers.internal:host-gateway

# Re-create container1
podman rm -f container1
podman run -d --name container1 -p 8081:80 --add-host=host.containers.internal:host-gateway nginx:alpine

# Go into container 1
podman exec -it container1 sh

# Try to reach server on host <--- Failed
curl http://$(hostname).local:8000

# Try to reach server on host <--- Failed
curl http://host.containers.internal:8000

# Try to reach server on host <--- Failed
curl http://localhost:8000

# Try to reach the other container <--- Worked
curl http://host.containers.internal:8082

3. Trying to add --network=host

This failed because the host network mounted is the network of the VM in which podman is running, not actually the host network.

# Re-create container1
podman rm -f container1
podman run -d --name container1 --network=host nginx:alpine

# Go into container 1
podman exec -it container1 sh

# Try to reach server on host <--- Failed
curl http://$(hostname).local:8000

# Try to reach server on host <--- Failed
curl http://hots.containers.internal:8000

# Try to reach server on host <--- Failed
curl http://localhost:8000

# Try to reach the other container <--- Worked
curl http://localhost:8082

Conclussion

This doesn't seem to be possible without relying on the user to edit directly Podman networking settings (for example this should work). It appears that we cannot do something from ToolHive UI/CLI or MCP Optimizer. I will leave this issue open to track what was tried and the potential workaround

[dmjb]

I wonder if a good workaround is to tell users to run ToolHive in podman as well.

EDIT: The problem is that the UI will always create a server running on the Windows host. It seems like we should either recommend that the user changes the config.

[value-added-korea]

Did this ever get "fixed?", I have been to the Discord discussion and there was a comment that it was "fixed", if this mcp version is out of date then please point this repo to the current releaes of it.

The host I am using is Ubuntu 25.x Kernel 6.17x Podman rootless, and docker rootless and there seems to be no joy and no method to verify the optimizer is "alive or dead"

[aponcedeleonch]

@value-added-korea Hello, unfortunately no, this never got a proper workaround. Note that this issue is Windows-specific. There are documented workarounds for the Ubuntu and Podman combination. The main blocker is this upstream issue: containers/podman#14933, which basically says there's no way for a Podman container to query or access an HTTP server running on the host machine. In our case, that host server would be ToolHive and the container would be the optimizer. This is due to the double virtualization layer that Podman uses on Windows in order to run Linux containers.